The Storage Tipping Point(storagemojo.com)
storagemojo.com
The Storage Tipping Point
http://storagemojo.com/2015/06/29/the-storage-tipping-point/
24 comments
Also worth noting is that IBM i had one of (if not probably the) first implementation of containers, in the form of LPARs.
True. Add that many "inventions" in virtualization field were in VM/370 in... the 70's. Including running itself on itself which has been celebrated in several news articles for different prototypes over the past few years. Except that was a production system. ;)
Add that IBM tried to kill VM multiple times, and wanted everyone using System/360 to standardize around OS/360.
"The Mythical Man-Month" is the story of OS/360. Hence VM still exists.
"The Mythical Man-Month" is the story of OS/360. Hence VM still exists.
Lol @ last line. Well-put.
The article says that log-structured apps and filesystems conflict with underlying SSDs, because those are doing similar things and the two don't play well together. But then it says that "products that already incorporate log structured I/O will have a definite advantage" in the future. How can they have an advantage if they're conflicting with the storage layer? Is this a contradiction, or am I missing something?
The paper [0] has some suggestions: an API (already existing and somewhat supported, called TRIM) to tell the disk that certain sectors are no longer in use, and atomic writes.
I guess I buy the claim that you're better off starting with a log-structured file system that will use new APIs to allocate and free chunks on the SSD, than to use a conventional file system that the SSD tries to optimize
[0] https://www.usenix.org/system/files/conference/inflow14/infl...
I guess I buy the claim that you're better off starting with a log-structured file system that will use new APIs to allocate and free chunks on the SSD, than to use a conventional file system that the SSD tries to optimize
[0] https://www.usenix.org/system/files/conference/inflow14/infl...
Yup, storage is going to change a lot. RAID is nearly dead. That means a whole lot of failure detection and correction needs to be reimplemented elsewhere in the stack. We'll also need to build storage layers that can deal with the crazy performance profile of SMR hard drives (at least, until SSD's are cheaper).
Wait till memresistors get here. That will basically eliminate the distinction between ram and disk.
https://en.wikipedia.org/wiki/Memristor#1808
"They can potentially be fashioned into non-volatile solid-state memory, which would allow greater data density than hard drives with access times similar to DRAM, replacing both components.[14] HP prototyped a crossbar latch memory that can fit 100 gigabits in a square centimeter,[8] and proposed a scalable 3D design (consisting of up to 1000 layers or 1 petabit per cm3).[66] "
https://en.wikipedia.org/wiki/Memristor#1808
"They can potentially be fashioned into non-volatile solid-state memory, which would allow greater data density than hard drives with access times similar to DRAM, replacing both components.[14] HP prototyped a crossbar latch memory that can fit 100 gigabits in a square centimeter,[8] and proposed a scalable 3D design (consisting of up to 1000 layers or 1 petabit per cm3).[66] "
HP indefinitely postponed (canceled) memristors recently.
https://news.ycombinator.com/item?id=9810324
They seem to be here.
They seem to be here.
I was under the impression that the storage model in the system/38 was part of what was carried forward into the as400. Not the capability model, as much as the lack of "files" because everything is just persistent objects/segments.
Is that not the case?
But I'm not really sure how this by itself solves the problems of modern SSD's. Problems caused by two things, the differing segment/page sizes between the upper layers (OS/filesystem/database/etc) and the need to rewrite static portions of the SSD in order to wear level the entire address space. (and to a lesser extent the need to update small portions in the middle of existing pieces of data).
Its the second part that seems to be forgotten by lots of people suggesting alternatives to the current storage stack. Whats really necessary is a higher level communications protocol between the flash and the filesystem. One that provides the metadata to the ssd about the sizes of individual "objects" so that they may be keep together as atomic units. That way the flash layer knows if a write is to a larger object, or a new object itself that can be managed separately.
Is that not the case?
But I'm not really sure how this by itself solves the problems of modern SSD's. Problems caused by two things, the differing segment/page sizes between the upper layers (OS/filesystem/database/etc) and the need to rewrite static portions of the SSD in order to wear level the entire address space. (and to a lesser extent the need to update small portions in the middle of existing pieces of data).
Its the second part that seems to be forgotten by lots of people suggesting alternatives to the current storage stack. Whats really necessary is a higher level communications protocol between the flash and the filesystem. One that provides the metadata to the ssd about the sizes of individual "objects" so that they may be keep together as atomic units. That way the flash layer knows if a write is to a larger object, or a new object itself that can be managed separately.
Good to see Robin Harris' article on the front page of HN (we live in the same small town and a mutual friend introduced us).
A little off topic from the article: I think the biggest tipping point is the decreasing cost of cloud storage because large companies like Google do so much of their time critical processing with data in RAM (spread over many servers) and these servers can re-purpose a lot of their disk I/O capability to service low cost cloud storage.
A little off topic from the article: I think the biggest tipping point is the decreasing cost of cloud storage because large companies like Google do so much of their time critical processing with data in RAM (spread over many servers) and these servers can re-purpose a lot of their disk I/O capability to service low cost cloud storage.
Can't believe I've seen two people from there on the same day on HN! Flabbergasted. Grew up there, currently living just up the rim.
Re: your comment, I'm curious to see how the I/O capacity of those servers could be distributed for cloud storage. I would imagine that even with 10Gbps if those machines are really cranking, they could have a difficult time provisioning network I/O for a storage service.
Re: your comment, I'm curious to see how the I/O capacity of those servers could be distributed for cloud storage. I would imagine that even with 10Gbps if those machines are really cranking, they could have a difficult time provisioning network I/O for a storage service.
Email me when you will be in Sedona and we can have coffee. I can contact Robin to see if he is available.
re: I/O capacity: I was commenting on how Google, etc. can provide such inexpensive cloud storage, not access speed.
re: I/O capacity: I was commenting on how Google, etc. can provide such inexpensive cloud storage, not access speed.
I think the storage company Nimble may have figured this out.
What makes you think they have? Do you have a blog post or statistical data that supports this notion? Please, share!
Not yet but the rep is taking us to the big game— a fully catered private box. I'm sure I'll hear about it there.
If they had truly solved this problem they wouldn't need to take you to "the big game" to do sales - you'd be taking them to "the big game" to be customers...
What is this "big game"? Are they taking you to a sporting event?
Sporting events seem like an inefficient place to transfer technical details. They're rather distracting.
Sporting events seem like an inefficient place to transfer technical details. They're rather distracting.
So their prices are for storage and the big game. Big sales seems so wasteful.
Nope. Nimble use a log storage format similar to that of Netapp and ZFS.
Their performance special source is using the minimum amount of flash to provide a performance boost over standard HDDs.
They also have very nice analytics.
They are good for VMware/virtualisation. For general file storage, unless they have changed recently, is pretty mediocre.
Their performance special source is using the minimum amount of flash to provide a performance boost over standard HDDs.
They also have very nice analytics.
They are good for VMware/virtualisation. For general file storage, unless they have changed recently, is pretty mediocre.
Not really, just an auto-tiering hybrid array, next generation storage subsystems i.e. those leveraging NVRAM are where we will likely stop talking blocks and start talking bytes just like we do to DRAM today. To get the greatest benefit of NVRAM the entire IO stack needs overhauling.
I thought they just run a rack full of managed SSDs?
I'd like to see FOSS do designs like that [minus the interface & API]. The labor advantage might lead to something pretty awesome. Meanwhile, they keep building complexity on top of UNIX architecture with a few stragglers on better OS architectures. Not much hope there. Fortunately, CHERI (capability) and SAFE (tags) are each building on mechanisms System/38 used for software reliability and security. An academic team or company might build something great on top of it. Still holding out some hope!
[1] http://homes.cs.washington.edu/~levy/capabook/Chapter8.pdf